RUC-NLPIR/Awesome-Long-Horizon-Agents

This GitHub repository is a curated collection of resources (papers, benchmarks, news, and more) about long-horizon AI agents — AI systems that can work through extended, multi-step tasks over minutes, hours, or even open-ended streams. The project accompanies a survey paper and aims to serve as a map of the field for researchers and developers.

What does it contain?

The repository is essentially a well‑organized list of 750+ papers and other resources, grouped by topic. Each entry typically links to the original paper or code. The README also includes:

• A survey paper (“Towards Long‑Horizon Agents: A Survey”) available on OpenReview and a preprint.
• News about updates and community features (e.g., a planned updated version).
• A table of contents that mirrors the structure of the survey.
• Figures that illustrate key ideas (e.g., the growth of task time horizons, the landscape of research).
• A citation file (BibTeX) so you can properly credit the project if you use it.

How is it organized?

The content is divided into major sections that guide you through the topic:

• Foundations — formalizes what “long‑horizon” means, with three nested levels (H1–H3) and the capabilities they demand.
• Evolution — traces the field through three stages: prompt engineering (2020–2023), context engineering (2023–2025), and runtime harnesses (2025–present).
• Harnesses — covers external engineering: loops, memory, tools, orchestration, verification, etc.
• Optimization — covers internal model adjustments: architecture, training data, fine‑tuning, reinforcement learning, self‑evolution.
• Applications — examples of long‑horizon agents in real‑world domains (software, information seeking, computer use, etc.).
• Benchmarks and Resources — evaluation tools and datasets.
• Frontiers — open problems and future directions.

This division reflects a central idea in the survey: long‑horizon ability comes from two interacting sides — a harness (external scaffolding like memory and tool use) and model optimization (internal training to handle long tasks). The README explains that these two sides co‑evolve: capabilities are first implemented in the harness, then possibly internalized into the model, and stronger models enable better harnesses.

Who is it for?

• Researchers wanting a structured overview of the field or a convenient bibliography.
• Developers building AI agents that need to operate over long time horizons.
• Students or enthusiasts looking for a guided introduction to the topic.

How can you use it?

• Browse the sections by clicking the links in the table of contents or scrolling the README.
• Read the survey – the paper explains the concepts and references many items in the list.
• Cite the project if you use it – a ready‑to‑use BibTeX entry is provided.
• Contribute – the repository explicitly welcomes pull requests, so you can add recent work.

Notable features

• Continuously updated – the README mentions a planned update in early November.
• Community‑driven – contributions are encouraged.
• Includes a website and social media links (X, Xiaohongshu, and WeChat coverage) for further updates.

In short, this project is a living, well‑structured repository that consolidates the key literature and resources on long‑horizon AI agents, making it easy to explore the topic from foundations to frontiers.